Why Not Use an Oracle When You Got One?

The primary goal of predictive modeling is to achieve high accuracy when the model is applied to novel data. For certain problems this requires the use of complex techniques like neural networks or ensembles, resulting in opaque models that are hard or impossible to interpret. For some domains this is unacceptable, since models need to be comprehensible. To achieve comprehensibility, accuracy is often sacrificed by using simpler techniques; a tradeoff termed the accuracy vs. comprehensibility tradeoff. Another, frequently studied, alternative is rule extraction; i.e. the activity where another, transparent, model is generated from the opaque model. In this paper it is argued that existing rule extraction algorithms do not use all information available, and typically should benefit from also using oracle data; i.e. test set instances, together with corresponding predictions from the opaque model. The experiments, using fifteen publicly available data sets, clearly show that rules extracted using either just oracle data or training data augmented with oracle data, will explain the predictions significantly better than rules extracted in the standard way; i.e. using training data only. Keywords—Rule extraction, neural networks, accuracy vs. comprehensibility tradeoff